|[eigen] D&C SVD ready to be used|
[ Thread Index |
| More lists.tuxfamily.org/eigen Archives
- To: eigen <eigen@xxxxxxxxxxxxxxxxxxx>
- Subject: [eigen] D&C SVD ready to be used
- From: Gael Guennebaud <gael.guennebaud@xxxxxxxxx>
- Date: Fri, 17 Oct 2014 18:04:45 +0200
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20120113; h=mime-version:from:date:message-id:subject:to:content-type; bh=nOJr4xdXsfftCS1r8SurmPgGdOcK3wIbFWn8W2bi4Mo=; b=PnA3nxlYefV2B+YUxIekTQ7e6EVC7wxp6g3SG9ywCAYSj888wnuUS/TBvZt4R4qWn0 hqsS7PqZfG2DMc64wrwdzuzqbuwJwLmCH9cnGWF1SeIf3TAQFqPuyE241mGqYPb6tGQR cPSUntaon/fzFKhlgUEro/5XXMvzfFh49T/v7tHtlwdFUOwUP1BM+02U3a123hCA1H/9 6jHt3Mk/ASIgm9w0LB9CL55rCE1HohcJP7TKWSsqqE9WCVY1txNud2tFbasKD+4NBKe1 pPPRrGdXx53rMBvbiT7Imvo1lwDnhnXPVWhmCN6kp+L/SzewJL9BMZL8UjZSCXn+Ih16 n5Gw==
during the last two months I've spent some time fixing the numerous numerical issues and bugs in BDCSVD. Unlike JacobiSVD for which it is enough to be able to compute a 2x2 SVD robustly to be good, the D&C algorithm is some order of magnitude more tricky!
Nevertheless, after tedious efforts, I'm glad that it finally passes both our unit tests (which I made much more tough) and Lapack's ones :) So I think it's time to move it to the official SVD module so that it will become part of the next 3.3 release. If you look at the code, you'll see a lot of ugly debugging statements which are disabled by default. Of course, I plan to clean them at some point, but since debugging BDCSVD appeared to be very very tricky and time consuming, I'd like to keep them during the stabilisation of the devel branch, just in case... Any objections?